Technical Sessions

As computer systems become ever-so complex to manage and optimize, various machine learning or data mining techniques have become popular in analyzing a large amount of system data. In this talk, I will share my limited experience and challenges we have encountered when exploring these techniques to solve system problems in our research projects and also commercial products.

Yuanyuan Zhou is currently a Qualcomm Chair Professor at the University of California, San Diego. Before UCSD, she was a tenured associate professor at University of Illinois at Urbana Champaign. Her research interests span the areas of operating systems, software engineering, system reliability and maintainability. She has co-founded three startups. Her recent startup, PatternInsight, has successfully deployed software quality assurance tools in many companies. In 2012, its Log Insight business line was acquired by VMware and now Log Insight is a VMware Product offering to its many data center customers for data center management. Dr. Zhou is an ACM Fellow and obtained her Ph.D. from Princeton. She has the great fortune of working with many talented students and colleagues.

Storage systems are often deployed in a tiered form to enable high performance and availability. These tiers utilize all possible volatile and non-volatile storage technologies, including DRAM, SSD, and HDD. The tradeoffs among their cost, features, and capabilities can make their effective integration into a single storage entity complex. Here, we propose an autonomic technique that learns user traffic patterns in a storage system over long time-scales to optimize user performance but also volume of completed system work. Our purpose is to multiplex as best as possible user workload with storage system features (e.g., voluminous internal system work) such that the latter is not starved but rather completed with minimal impact on user performance. Key to achieving the above is to use an autonomic learning engine to predict when the user workload intensity increases/ decreases and then proactively stop/start bulky internal system work. Being proactive allows the system to effectively bring into the fast tier the active user working set just-in-time and right before it is needed most, i.e., when user traffic suddenly peaks. We illustrate the effectiveness of this mechanism by using both trace driven simulations from production systems as well experiments on a real testbed.

Control theory provides solid foundations for developing reliable and scalable feedback control for software systems. Although, feedback controllers have been acknowledged to efficiently solve common classes of problems, their adoption by state-of-the-art approaches for designing self-adaptation in legacy software systems remains limited and at best consists in ad hoc integrations, which are usually engineered manually.

In this paper, we revisit the Znn.com case study and we present an alternative implementation based on classical feedback controllers. We show how these controllers can be easily integrated into software systems through control theory centric architecture models and domainspecific modeling support. We also provide an assessment of the resulting properties, quality attributes and limitations.

Motivated by the growing popularity of database-as-aservice clouds, this paper presents ShuttleDB, a holistic approach enabling flexible, automated elasticity of database tenants in the cloud. We first propose a database-aware live migration and replication method designed to work with off-the-shelf databases without any database engine modifications. We then combine these database-aware techniques with VM-level mechanisms to implement a flexible elasticity approach that can achieve efficient scale up, scale out, or scale back for diverse tenants with fluctuating workloads. Our experimental evaluation of the ShuttleDB prototype shows that by applying migration and replication techniques at the tenant level, automated elasticity can be achieved both intra- and inter-datacenter in a database agnostic way. We further show that ShuttleDB can reduce the time and data transfer needed for elasticity by 80% or more compared to tenant-oblivious approaches.

Applications with a dynamic workload demand need access to a flexible infrastructure to meet performance guarantees and minimize resource costs. While cloud computing provides the elasticity to scale the infrastructure on demand, cloud service providers lack control and visibility of user space applications, making it difficult to accurately scale the underlying infrastructure. Thus, the burden of scaling falls on the user.

In this paper, we propose a new cloud service, Dependable Compute Cloud (DC2), that automatically scales the infrastructure to meet the user-specified performance requirements. DC2 employs Kalman filtering to automatically learn the (possibly changing) system parameters for each application, allowing it to proactively scale the infrastructure to meet performance guarantees. DC2 is designed for the cloud - it is application-agnostic and does not require any offline application profiling or benchmarking. Our implementation results on OpenStack using a multi-tier application under a range of workload traces demonstrate the robustness and superiority of DC2 over existing rule-based approaches.

We present a novel application of deep learning in networking. The envisioned system can learn the original flow characteristics such as a burst size and inter-burst gaps conceived at the source from packet sampling done at a receiverWi-Fi node. This problem is challenging because CSMA introduces complex, irregular alterations to the origin pattern of the flow in the presence of competing flows. Our approach is semi-supervised learning. We first work through multiple layers of feature extraction and subsampling from unlabeled flow measurements.We use a feature extractor based on sparse coding and dictionary learning, and our subsampler performs overlapping max pooling. Given the layers of learned feature mapping, we train SVM classifiers with deep feature representation resulted at the top layer. The proposed scheme has been evaluated empirically in a custom wireless simulator and OPNET. The results are promising that we achieve superior classification performance over ARMAX, Naïve Bayes classifiers, and Gaussian mixture models optimized by the EM algorithm.

A controller is implemented to manage memory as an elastic resource similar to computing cycles for Java applications. The controller actively arbitrates constrained memory between collocated JVMs in response to demand. A key aspect of the work is that JVM metrics are used as proxies for application KPIs so that application performance instrumentation and modeling are not required. A metric corresponding to the allocation rate of memory is derived from the JVM metrics and established as the measure of application performance and is used as the effective feedback mechanism to the controller. The controller is based on a fair share policy in which memory is distributed to equalize the marginal performance value to all JVMs. The design is tested for effectiveness and stability using the suite of SPECjvm2008 and SPECjbb2005 benchmarks.

Processing ever-increasing amounts of information and providing a meaningful analysis of large datasets (Big Data) has become a significant computing challenge in the Enterprise environment. New tools, frameworks, and systems have been proposed for Big Data processing. They target a variety of data (everything from business transactions to sensor data to tweets) and aim to offer new useful insights via advanced real-time analytics and/or batch-driven data analysis. The common theme of these underlying systems is that they represent a scale-out approach on commodity machines. Using a MapReduce framework I will present and analyze challenges in performance management of such systems. I will talk about the community and enterprise efforts to design unified and/or integrated data processing frameworks that aim to simplify application development and enhance data analytics. Finally, I will discuss hardware and resource usage patterns imposed by modern and emerging scale-out applications and their possible impact on the future system design.

Dr. Ludmila Cherkasova is a principal scientist in the Systems Research Lab, at Hewlett-Packard Labs, Palo Alto, USA. Her research interests include the analysis, design, and management of concurrent and distributed systems (such as emerging systems for Big Data processing, internet and media applications, virtualized environments, and next generation data centers). She has authored over 100 referred publications and more than 70 patent applications. Over the years she has mentored and co-advised more than 15 PhD students/interns. She has served as a Track/PC co-chair of 6 International conferences. She is an ACM Distinguished Scientist and recognized by the Certificate of Appreciation from the IEEE Computer Society. She has earned 6 Best Paper awards. Her most recent works were on the design of performance analysis and optimization techniques for MapReduce environments.

Efficient namespace metadata management is increasingly important as next-generation storage systems are designed for peta and exascales. New schemes have been proposed; however, their evaluation has been insufficient due to a lack of an appropriate namespace metadata benchmark. We describe MimesisBench, a novel namespace metadata benchmark for next-generation storage systems, and demonstrate its usefulness through a study of the scalability and performance of the Hadoop Distributed File System (HDFS).

We demonstrate an enterprise Dynamic Thresholding System for data-agnostic management of monitoring flows. The dynamic thresholding based on data historical behavior enables adaptive and more accurate control of business environments compared to static thresholding. We manifest the main blocks of a complex analytical engine that is implemented in VMware vCenter Operations Manager as a principal foundation of the company’s data-driven anomaly detection.

There is a movement in many practical applications of Cyber-Physical Systems to push processing to the edge. This is particularly important were the CPS is carrying out monitoring and control, where the latency between the decision making and control message reception should be minimal. However, CPS are limited by the capabilities of the typically battery powered low resourced devices. In this paper we present a self-adaptive scheme that both reduces the amount of resources required to store high sample rate data at the edge and at the same time carries out initial data analytics. Using out Smart Water datasets, plus a selection from other real world CPS applications, we show that our algorithm reduces computation by 98%; data volumes by 55%; while requiring only 11KB of memory at runtime (including the compression algorithm). In addition we show that our system supports self-tuning and automatic reconfiguration which means that manual tuning is alleviated and the scheme can be both applied to any kind of raw data automatically and is able self-optimize as the nature of the incoming data changes over time.

Gait signals detectable by sensors on ubiquitous personal devices such as smartphones can reveal characteristics unique to each individual, and thereby offer a new approach to recognizing users. Conventional pattern matching approaches use inner-product based distance measures which are not robust to common variations in time-series analysis (e.g., shifts and stretching). This is unfortunate given that it is well understood that capturing such variations is paramount for model performance. This work shows how machine learning methods which encode gait signals into a feature space based on a dictionary can use convolution and Dynamic TimeWarping (DTW) similarity measures to improve classification accuracy in a variety of situations common to gait recognition. We also show that data augmentation is crucial in gait recognition, as diverse training data in practical applications is very limited. We validate the effectiveness of these methods empirically, and demonstrate the identification of user gait patterns where shift and stretch variations in measurements are substantial. We present a new gait dataset that contains a complete representation of the variations that can be expected in real-world recognition scenarios. We compare our techniques against the current state of the art gait period detection and normalization schemes on our dataset and show improved classification accuracy under all experimental scenarios.

Today, big and small organizations alike collect huge amounts of data, and they do so with one goal in mind: extract "value" through sophisticated exploratory analysis, and use it as the basis to make decisions as varied as personalized treatment and ad targeting. Unfortunately, existing data analytics tools are slow in answering queries, as they typically require to sift through huge amounts of data stored on disk, and are even less suitable for complex computations, such as machine learning algorithms. These limitations leave the potential of extracting value of big data unfulfilled.

To address this challenge, we are developing Berkeley Data Analytics Stack (BDAS), an open source data analytics stack that provides interactive response times for complex computations on massive data. To achieve this goal, BDAS supports efficient, large-scale in-memory data processing, and allows users and applications to trade between query accuracy, time, and cost. In this talk, I'll present the architecture, challenges, results, and our experience with developing BDAS, with a focus on Apache Spark, an in-memory cluster computing engine that provides support for a variety of workloads, including batch, streaming, and iterative computations. In a relatively short time, Spark has become the most active big data project in the open source community, and is already being used by over one hundred of companies and research institutions.

Ion Stoica is a Professor in the EECS Department at University of California at Berkeley. He received his PhD from Carnegie Mellon University in 2000. He does research on cloud computing and networked computer systems. Past work includes the Dynamic Packet State (DPS), Chord DHT, Internet Indirection Infrastructure (i3), declarative networks, replay-debugging, and multi-layer tracing in distributed systems. His current research focuses on resource management and scheduling for data centers, cluster computing frameworks, and network architectures. He is an ACM Fellow and has received numerous awards, including the SIGCOMM Test of Time Award (2011), and the ACM doctoral dissertation award (2001). In 2006, he co-founded Conviva, a startup to commercialize technologies for large scale video distribution, and in 2013, he co-founded Databricks as startup to commercialize technologies for Big Data processing.

Cloud computing provides an attractive computing paradigm in which computational resources are rented on-demand to users with zero capital and maintenance costs. Cloud providers offer different pricing options to meet computing requirements of a wide variety of applications. An attractive option for batch computing is spot-instances, which allows users to place bids for spare computing instances and rent them at a (often) substantially lower price compared to the fixed on-demand price. However, this raises three main challenges for users: how many instances to rent at any time? what type (on-demand, spot, or both)? and what bid value to use for spot instances? In particular, renting on-demand risks high costs while renting spot instances risks job interruption and delayed completion when the spot market price exceeds the bid. This paper introduces an online learning algorithm for resource allocation to address this fundamental tradeoff between computation cost and performance. Our algorithm dynamically adapts resource allocation by learning from its performance on prior job executions while incorporating history of spot prices and workload characteristics. We provide theoretical bounds on its performance and prove that the average regret of our approach (compared to the best policy in hindsight) vanishes to zero with time. Evaluation on traces from a large datacenter cluster shows that our algorithm outperforms greedy allocation heuristics and quickly converges to a small set of best performing policies.

Data centers are promising participants in demand response programs (i.e., reducing a large electricity demand upon utility’s request), making power grid more stable and sustainable. In this paper, we focus on enabling colocation data center demand response. Colocation is an integral yet unique segment of data center industry, where multiple tenants house their servers in one shared facility. Nonetheless, differing from owner-operated data centers (e.g., Google), colocation data center suffers from “split incentive”: colocation operator desires demand response for financial incentives but has no control over tenants’ servers, while tenants who own the servers may not desire demand response due to lack of incentives. To break “split incentive”, we propose a first-of-its-kind incentive mechanism, called iCODE (incentivizing COlocation tenants for DEmand response), based on reverse auction: tenants, who voluntarily submit energy reduction bids to colocation operator, will be financially rewarded if their bids are accepted. We formally model how each tenant decides its bids and how colocation operator decides winning bids. We perform a trace-based simulation to evaluate iCODE. We show that iCODE can reduce colocation energy consumption by over 50% during demand response periods, unleashing the potential of colocation demand response.

Transactional Memory was recently integrated in Intel processors under the name TSX.We show that its performance can be significantly affected by the configuration of its interplay with the software-based fallback: in fact, there does not seem to exist a single configuration that can perform best independently of the application and workload. We address this challenge by introducing an innovative self-tuning approach that exploits lightweight reinforcement learning techniques to identify the optimal TSX configuration in a workload-oblivious manner, i.e. not requiring any off-line/a-priori sampling of the application’s workload. To achieve total transparency for the programmer, we integrated the proposed algorithm in the GCC compiler. Our evaluation shows improvements up to 2× over state of the art approaches, while remaining within 5% from the performance achievable using optimal static configurations.

This paper presents an autonomic resource management solution that looks at the non-functional qualities of a Web-based system, as well as at the characteristics of the infrastructural resources it uses. It exploits a detailed performance model of both these aspects to increase the efficiency of resource allocation. The solution is evaluated on an auction and shopping benchmark web site, and compared to a baseline approach and to an existing solution from literature. Results show that, by jointly taking into account the different software and hardware facets of our application, we can reduce the amount of resources allocated by up to 42.5% compared with an existing work from the literature.

Data centers are seeking more efficient cooling techniques to reduce their operating expenses, because cooling can account for 30-40% of the power consumption of a data center. Recently, liquid cooling has emerged as a promising alternative to traditional air cooling, because it can help eliminate undesired air recirculation. Another emerging technology is free air cooling, which saves chiller power by utilizing outside cold air for cooling. Some existing data centers have already started to adopt both liquid and free air cooling techniques for significantly improved cooling efficiency and more data centers are expected to follow.

In this paper, we propose SmartCool, a power optimization scheme that effectively coordinates different cooling techniques and dynamically manages workload allocation for jointly optimized cooling and server power. In sharp contrast to the existing work that addresses different cooling techniques in an isolated manner, SmartCool systematically formulates the integration of different cooling systems as a constrained optimization problem. Furthermore, since geo-distributed data centers have different ambient temperatures, SmartCool dynamically dispatches the incoming requests among a network of data centers with heterogeneous cooling systems to best leverage the high efficiency of free cooling. A light-weight heuristic algorithm is proposed to achieve a near-optimal solution with a low run-time overhead. We evaluate SmartCool both in simulation and on a hardware testbed. The results show that SmartCool outperforms two state-of-the-art baselines by having a 38% more power savings.

This paper presents WattValet, an efficient solution to reduce data center peak power consumption by using heterogeneous energy storage. We henceforth call an energy storage device, a battery, with the understanding that the discussion applies to other devices as well such as pumped hydraulic and thermal systems. Previous work on energy storage management in data centers often ignores or underestimates their degree of heterogeneity. Even if batteries used in a data center are of the same model and purchased at the same time, differences in storing temperature and humidity, as well as discharging cycles and depth, gradually drive their characteristics apart. We show that differences in battery characteristics, such as discharge rates, if not fully accounted for, can lead to significantly suboptimal power caps. A new algorithm, called WattValet, is described that reduces peak power consumption by efficiently exploiting heterogeneity. Evaluation using Wikipedia traces shows that the power cap generated by WattValet is within 2% of the optimal solution, whereas WattValet finishes the computation orders of magnitude faster than the optimal solution even in small-scale experiments.